Persistency of priors-induced bias in decision

Original Research Article

published: 08 March 2011 doi: 10.3389/fnins.2011.00029

Persistency of priors-induced bias in decision behavior and the fMRI signal Kathleen A. Hansen1*, Sarah F. Hillenbrand 2 and Leslie G. Ungerleider1 Laboratory of Brain and Cognition, National Institute of Mental Health, National Institutes of Health, Bethesda, MA, USA Helen Wills Neuroscience Institute, University of California, Berkeley, CA, USA

1 2

Edited by: Paul Glimcher, New York University, USA Reviewed by: Christopher Summerfield, Oxford University, USA Ifat Levy, Yale University School of Medicine, USA *Correspondence: Kathleen A. Hansen, Laboratory of Brain and Cognition, National Institute of Mental Health, National Institutes of Health, Building 10 Room 4C104, Bethesda, MD 20892, USA. e-mail: [email protected]

It is well known that people take advantage of prior knowledge to bias decisions. To investigate this phenomenon behaviorally and in the brain, we acquired fMRI data while human subjects viewed ambiguous abstract shapes and decided whether a shape was of Category A (smoother) or B (bumpier). The decision was made in the context of one of two prior knowledge cues, 80/20 and 50/50. The 80/20 cue indicated that upcoming shapes had an 80% probability of being of one category, e.g., B, and a 20% probability of being of the other. The 50/50 cue indicated that upcoming shapes had an equal probability of being of either category. The ideal observer would bias decisions in favor of the indicated alternative at 80/20 and show zero bias at 50/50. We found that subjects did bias their decisions in the predicted direction at 80/20 but did not show zero bias at 50/50. Instead, at 50/50 the subjects retained biases of the same sign as their 80/20 biases, though of diminished magnitude. The signature of a persistent though diminished bias at 50/50 was also evident in fMRI data from frontal and parietal regions previously implicated in decision-making. As a control, we acquired fMRI data from naïve subjects who experienced only the 50/50 stimulus distributions during both the pre-scan training and the fMRI experiment. The behavioral and fMRI data from the naïve subjects reflected decision biases closer to those of the ideal observer than those of the prior knowledge subjects at 50/50. The results indicate that practice making decisions in the context of non-equal prior probabilities biases decisions made later when prior probabilities are equal. This finding may be related to the “anchoring and adjustment” strategy described in the psychology, economics, and marketing literatures, in which subjects adjust a first approximation response – the “anchor” – based on additional information, typically applying insufficient adjustment relative to the ideal observer. Keywords: choice, experience, expectation

INTRODUCTION When making decisions, people take advantage of available prior knowledge to bias their choices (Green and Swets, 1966). This common-sense behavior increases the chance that decisions will be correct. In the laboratory, researchers study the effects of prior knowledge on decision bias by asking subjects to make choices in the context of two or more prior knowledge conditions. For example, consider a prior knowledge condition indicating that Alternative 1 has an 80% and Alternative 2 has a 20% chance of being the correct choice; we will call this an 80/20 prior knowledge condition. In many experiments (Green and Swets, 1966; Healy and Kubovy, 1978, 1981; Maddox, 2002), subjects trained and tested on an 80/20 prior knowledge condition are also trained and tested on the inverse condition: 20/80, in which Alternative 1 has an 20% and Alternative 2 has an 80% chance of being the correct choice. In some cases the 50/50 condition, in which each alternative has a 50% chance of being the correct choice, is also tested. Under such experimental conditions, the performance of human subjects approximates that of the ideal observer, who would bias decisions in favor of the indicated alternatives at 80/20 and 20/80 and exhibit zero bias at 50/50 (Green and Swets, 1966; Healy and Kubovy, 1978, 1981; Maddox, 2002).

www.frontiersin.org

Inverse prior knowledge conditions are convenient for counterbalancing experimental factors in the laboratory. In the real-world, however, inverse prior knowledge conditions are rarely experienced during a time period as short as that of a typical experiment. In a more common real-world scenario, a certain prior knowledge condition can be relevant to a decision at one time, indicating that a bias is then appropriate, but cease to be relevant to a decision at a later date. People often fail to adopt the appropriate bias of zero in the later decision, presumably because they have difficulty ignoring the previously learned but no longer relevant prior knowledge. This phenomenon is familiar to us all. In fact, although decision researchers use the word bias to refer to an optimizable quantity, the common English usage connotes an undesirable influence that ideally should be set aside. Thus, the typical laboratory approach of inverting prior knowledge conditions within subjects does not adequately reflect real-world constraints. To address this problem experimentally, we probed the behavioral and fMRI responses of human subjects viewing ambiguous abstract shapes and deciding whether a shape was of Category A (smoother) or B (bumpier). The decision was made in the context of one of two prior knowledge cues, 80/20 and 50/50. The 80/20 cue meant that upcoming shapes had an 80% probability of being of

March 2011 | Volume 5 | Article 29 | 1

Hansen et al.

Persistency of priors-induced bias

one category, e.g., B, and a 20% probability of being of the other; we refer to the 80 and 20% categories as indicated and contraindicated respectively. The 50/50 cue meant that upcoming shapes had an equal probability of being of either category. Subjects learned the meaning of the cues in pre-scan training runs. During training, the 80/20 and 50/50 cues were accompanied by 80/20 and 50/50 target distributions, respectively; the training distributions were created by manipulating the prior probability of occurrence of the physical targets themselves, rather than changing the category boundary. No subject experienced inverse prior knowledge conditions; for example, a subject who learned that 80/20 indicated Category A never had to relearn the task with a 20/80 cue contraindicating Category A. We found that subjects’ decisions made in the context of both the 80/20 cue and the 50/50 cue were biased in the direction indicated by the 80/20 cue. In the 50/50 condition, the magnitude of the bias was diminished relative to the 80/20 condition, but failed to reach the zero bias predicted for the ideal observer. The persistent bias suggested that even when the chance of either target type was equal, the targets were processed at some level by the prior knowledge subjects as indicated or contraindicated. Therefore, we predicted that, in some brain areas, differences in fMRI activation elicited by indicated vs. contraindicated targets in the 80/20 runs would be persistent, though perhaps diminished, in the 50/50 runs. This hypothesis found confirmation in fMRI data from frontal and parietal regions previously implicated in decision-making. As a control, we acquired fMRI data from naïve subjects who experienced only the 50/50 stimulus distributions during both the pre-scan training and the fMRI experiment. The behavioral and fMRI data from these naïve subjects reflected decision biases closer to those of the ideal observer than those of the prior knowledge subjects at 50/50. These findings have important implications for understanding decisionmaking under ambiguity in real-world conditions.

Figure 1 | Target distributions and visual appearance. (A) The distributions of Category A and B shapes shown during all fMRI runs were Gaussian and overlapping. Curvature levels between 9 and 17% of the radius were ambiguous, i.e., shapes of these curvature levels could be of either category. (B) The visual appearance of all curvature levels used. Neighboring shapes were difficult or impossible to discriminate from one another, preventing subjects from using a visual memorization strategy to perform the task. (C) In the experiment, the shapes were jittered in size, position, and orientation, as in the examples shown here.

MATERIALS AND METHODS Participants

In this study, we acquired fMRI and behavioral data from 58 subjects, all of whom provided informed consent before the experiment. All procedures were approved by the National Institute of Mental Health Institutional Review Board. All subjects were righthanded and had normal or corrected-to-normal vision. Here we present the data from 45 subjects (22 male) of mean age 25 years (range 20–41). Data from the remaining subjects were excluded because of a report that uncomfortably dry eyes prevented the subject from focusing on the stimuli, a broken shim coil, unacceptably low estimates of d′ or patterns of random button presses that led to poor fits to psychometric functions.

distorted circle stimuli were created in MATLAB (Version 7.31) according to and adapted from equations from Wilkinson et al. (1998). The shape contour of each stimulus, r(q), was created by sinusoidally modulating the radius of a circle: r (θ) = rmean (1 + A sin(ωθ + φ))

where r and q (in radians) are the polar coordinates of the contour, rmean is its mean radius and A, w, f are, respectively, the amplitude (expressed as a proportion of the radius), radial frequency, and phase of the modulation. Setting A to 0 defines a perfect circle. The cross-sectional profile of each stimulus, c, was modified by blurring the shape contour exponentially:

Stimuli and task

Targets were distorted circles (Wilkinson et al., 1998) whose sinusoidal modulation ranged linearly from 4 to 22% of the mean radius, with step size 0.5%. The smoothest target was defined as the Category A prototype, and the bumpiest as the Category B prototype (Figure 1). Distributions of Category A and B shapes were Gaussian and overlapping (Healy and Kubovy, 1981; Maddox, 2002). The overlapping distributions made intermediate targets ambiguous, so that the targets alone would not contain sufficient information for subjects to classify them with perfect accuracy. The

Frontiers in Neuroscience | Decision Neuroscience

(1)

2

c = e− ( r − r ( θ) / σ)

(2)

where r is the set of all distances between the central point and the image edge, r(q) is as defined in Eq. 1, and s determines the peak spatial frequency of the output image (peak spatial frequency = √2 / ps). www.mathworks.com

1


Hansen et al.

The color of the distorted circles was converted to black and the background was converted to gray. Stimuli were presented with the Presentation software (Version 10.22) and projected onto a translucent screen placed at the foot of the scanner bed. Subjects viewed a reflection of the back-projected stimuli. The task (Figure 2) was to decide whether a shape was Category A or B. The shapes were presented one at a time with random sizes, orientations, and locations to encourage the use of stimulus shape to make decisions and to prevent subjects from relying on retinotopic location or spatial attention in order to perform well. No part of any shape subtended more than two radial degrees, and the location of the fixation cross was inside each shape. Before each shape a cue was presented; the same cue was used throughout each run. To ensure that the subject did not forget the prior knowledge condition during the run, the cue was repeated at the beginning of each trial. Before entering the scanner, 22 subjects underwent behavioral training that included explicit prior knowledge cues, 80/20 and 50/50. The indicated target category – that is, the category indicated by 80 in the 80/20 training runs – was A for 8 subjects and B for 14 subjects. In the training, two 80/20 and two 50/50 runs were interleaved. For each subject, the order was 50/50 run 1, 80/20 run 1, 50/50 run 2, 80/20 run 2. The 80/20 training runs were comprised of 80% indicated (i.e., having curvature smoother than the mean sinusoidal modulation of 13% if the indicated category was A, or bumpier than 13% if the indicated category was B) and 20% contraindicated targets. The 50/50 runs were comprised of 50% of each target type. Thus, during training, the explicit prior knowledge cues reflected the implicit prior probability distributions of the targets. Subjects received feedback after each training trial. These 22 subjects were informed explicitly that the target distributions were 80/20 and 50/50, and their understanding of this concept was confirmed by their answers to questions during pre-training instruction. For these subjects, the scanning runs differed from training runs in three respects. First, all scanning runs were comprised of 50% indicated and 50% contraindicated targets, such that the targets in each 80/20 run were identical to the targets in a 50/50 run. This control ensured that differences between prior 2

www.neurobs.com


knowledge conditions could be attributed only to the cue and not to stimulation differences. Second, subjects did not receive feedback during scanning. Third, one-third of the trials in each scanning run were catch trials, in which a blank screen took the place of the target and subjects were instructed to make no response. The inclusion of catch trials permitted us to obtain estimates of activity during decisions vs. catch trials within each priors cue condition. The remaining 23 subjects underwent pre-scan behavioral training at the 50/50 distribution only and experienced the sham cue (OO/OO) during both the training and the fMRI experiment. These naïve subjects were never exposed to the 80/20 distributions experienced during training by the prior knowledge subjects, and were not informed explicitly that the underlying distributions were always 50/50. In other respects, the instructions, training, and fMRI experiment were identical for the naïve subjects and the prior knowledge subjects. The order of trial types (Category A target, Category B target or catch trial) for the scanning runs was determined by assigning each run a different ternary m-sequence. M-sequences are efficient in terms of signal per time, especially for relatively short scan durations, and are exactly counterbalanced over time, minimizing any uncontrolled adaptation or expectation effects (Sutter, 2001; Buracˇas and Boynton, 2002). M-sequences were generated using code written by G. Buracˇas (Buracˇas and Boynton, 2002). Each runlength m-sequence was length 34 − 1 = 80, consisting of 27 Category A stimulus trials, 27 Category B stimulus trials, and 26 catch trials; thus 33% of the trials were catch trials. Each trial lasted 2.5 s. A blank grayscale screen was shown for 10 s at the beginning of each run to allow the magnetic field to reach equilibrium and for 12.5 s at the end of each run to allow for the delay in the hemodynamic response. The cue was 50/50 on six runs and 80/20 on six runs, with the cue type alternating pseudorandomly from run to run. Imaging data acquisition and preprocessing

All MRI data were collected on a GE 3-Tesla scanner with a GE whole-head 8-channel coil. For fMRI we used an echo-planar imaging (EPI) sequence with repetition time (TR) = 2.5 s per shot (=2.5 s per acquired brain volume), echo time (TE) = 30 ms, field of view 22 cm by 22 cm, resolution 64 × 64 voxels per slice (in-plane voxel size 3.4 mm × 3.4 mm), and slice thickness 3.0 mm. Each fMRI

Figure 2 | Trial structure during training and scanning. For decision trials, subjects were instructed to decide whether each target was of Category A or B. For catch trials, subjects were instructed to continue fixating and await the next trial. A small fixation dot was present during the stimulus epoch of each trial and during the delay and response epochs of both decision and catch trials.

www.frontiersin.org


Hansen et al.


brain volume consisted of 38 axial slices. For anatomical images we used an magnetization prepared rapid acquisition gradient echo (MP-RAGE) sequence with field of view 24 cm by 24 cm, 128 locations per slab and slice thickness 1.2 mm. Unless otherwise noted, preprocessing and subsequent analysis of the MRI data was performed with the AFNI software package (Cox, 1996; Cox and Hyde, 1997). Italics indicate AFNI function names. The first four brain volumes of every fMRI run were removed and brain volumes were time-shifted to account for the acquisition time of each slice. Data from each run were registered and motion-corrected using 3dvolreg. Each subject’s T1-weighted anatomical dataset was warped via 12-parameter affine transform to a single template volume (the N27 “Colin” brain) in Talairach space using @auto_tlrc.

by 3dvolreg. The GLM analysis was performed using 3dDeconvolve. Outputs were voxelwise beta weights representing the percent signal change vs. baseline attributable to each regressor. Signal variability attributable to head motion estimates was assigned to the baseline. For each subject, the ROIs derived from the contraindicated vs. indicated analysis on the 80/20 data were converted to individual brain space. The betas corresponding to each subject’s fMRI responses to each of the nine curvature bins at 80/20 and 50/50, respectively, were sampled from and averaged within each individual ROI. The grand means and SE across subjects were calculated for each bin and prior knowledge condition, and the results were plotted as tuning curves across the dimension of curvature bins.

ROI identification

RESULTS

To identify regions of interest (ROIs), we first estimated fMRI responses in the 80/20 runs to the presentation of targets indicated and contraindicated by the 80/20 cue. Two sequences of 0s and 1s, where the 1s represented indicated and contraindicated targets respectively, were convolved with a model hemodynamic function using waver to create the regressors for the analysis. Other inputs to the GLM were the estimates of head motion produced by 3dvolreg. The GLM analysis was performed using 3dDeconvolve. Outputs were voxelwise beta weights representing the percent signal change vs. baseline attributable to each regressor. Signal variability attributable to head motion estimates was assigned to the baseline. A random effects analysis (random effect of subject) was performed on the betas produced by the individual GLMs. using 3dAnova2 to calculate the mean responses to indicated and contraindicated targets and to obtain indicated vs. contraindicated differences. From the group analysis results, a mask was derived identifying voxels where indicated vs. contraindicated differences, as well as either the indicated or contraindicated mean activity levels, exceeded uncorrected p